Instantaneous tracking of earthquake growth with elastogravity signals

The study addresses the long-standing challenge of rapidly and accurately estimating the size and evolution of very large (M > 8) earthquakes for effective earthquake and tsunami early warning. Conventional seismic-based EEWS suffer from magnitude saturation and latency because they rely on the first seconds of P-waves, which contain insufficient information to distinguish very large events early and are limited by wave propagation speed. Geodetic/GNSS-based approaches alleviate saturation but still face uncertainties, subjective preprocessing choices, and latency, and their apparent rapid success has been questioned as potentially arising from prior constraints. The authors propose leveraging prompt elastogravity signals (PEGS), which propagate effectively at the speed of light and are sensitive to the evolving moment release, to enable instantaneous tracking of earthquake growth once the event reaches a certain magnitude threshold. They develop a deep learning framework, PEGSNet, to estimate location and track time-dependent magnitude Mw(t) using pre-P-wave broadband seismograms from Japan.

Prior work established the potential of PEGS, observed prior to direct seismic waves, to inform earthquake source properties and overcome latency constraints in EEWS, with detectability suggested for events M > 8. Traditional EEWS based on P-wave features saturate for megathrust events; for example, JMA underestimated the 2011 Mw 9.0 Tohoku-Oki event to ~8.1. Finite-fault EEWS and real-time GNSS approaches have improved performance but are sensitive to data selection, preprocessing choices, and priors, and still incur latency due to seismic wave speeds. Recent deep learning applications in seismology have advanced detection, picking, and source characterization, including GNSS-based deep models, but these still require assumptions (e.g., slip distributions) and are constrained by P-wave propagation speed. The literature thus motivates exploring PEGS with deep learning for operational early warning without magnitude saturation and with minimal latency.

Study region: Japanese subduction zone, leveraging dense regional broadband networks (F-Net and other stations via IRIS). Training data: 500,000 synthetic PEGS waveforms generated by a normal-mode summation code for point-source ruptures placed along 1,400 discrete locations on the megathrust guided by Slab2.0 geometry. Source parameters: magnitudes uniformly sampled from Mw 5.5–10; rake angles from a Gaussian (μ = 90°, σ = 10°); strike and dip from Slab2.0. Source time functions (STFs): randomly generated to mimic empirical laws/statistical observations, allowing realistic variability beyond simple triangles. Waveforms: three components (E, N, Z) for each station; empirical noise recorded at each station added to synthetics. Preprocessing: decimation to 1 Hz; bandpass filtering between 2 mHz (two-pole causal Butterworth high-pass) and 30 mHz (six-pole causal Butterworth low-pass); clipping at ±10 nm s^2 and scaling by the same value; traces stored as 700 s long, centered at origin time. Input formatting: multi-station inputs arranged as image-like tensors (stations sorted by longitude) with three channels (components). To isolate PEGS, amplitudes for t ≥ Tp (P-wave arrival time at each station) are set to zero; P-wave arrival times are assumed known via standard triggering. Labels: latitude φ, longitude λ, and time-integrated STF converted to time-dependent Mw(t). Training strategy: during training, a start time T1 is randomly selected; 315 s windows from T1 to T1+315 s are used as inputs, with label Mw(T2) at the end of the window. Architecture: a LeNet-like convolutional neural network (convolutional layers followed by fully connected layers). Loss and selection: Huber loss minimized over Mw(T2), φ, λ; best model chosen by lowest validation error. Data splits: 70% training, 20% validation, 10% test with strict separation of STFs and noise records among splits. Evaluation protocol: simulate real time by sliding a 315 s window forward in 1 s steps, predicting Mw at each second to reconstruct Mw(t). Location performance (dependent on P arrivals) shows ~25–30 km errors starting ~50 s after origin (details in Extended Data). Real-data test: retrospective playback of the 2011 Mw 9.0 Tohoku-Oki earthquake with identical preprocessing, making second-by-second Mw(t) predictions using only the preceding 315 s data window.

• On synthetic tests, PEGSNet tracks Mw(t) for events with final Mw > 8.6 starting ~40 s after origin with accuracy > 90% (successful predictions within ±0.4 Mw units) and average errors < 0.25 Mw units.
• For final Mw between 8.2 and 8.6, early tracking is difficult; only final Mw is estimated after ~150 s with 60–70% accuracy and errors > 0.25. Predicted Mw below ~8.2 is poorly constrained; conservative sensitivity lower limit Mw ≈ 8.3. Under favorable low-noise conditions (network noise σ ≲ 0.5 nm s^2), sensitivity improves to ~Mw 7.9–8.0.
• Predictions for true Mw 9.0 ± 0.05 events are underestimated in the first 30–40 s (due to sensitivity threshold and P-wave masking at near-source stations). Noise-free synthetics indicate ≥15 s after origin are needed for reliable Mw estimation.
• From ~40 s onward, PEGSNet reconstructs Mw(t) without time shift (instantaneous tracking) relative to the true STF, leveraging PEGS’ speed-of-light information.
• Retrospective Tohoku-Oki test: after ~55 s, PEGSNet is consistently closer to the ‘true’ STF than JMA, FinDer2, and GNSS-based BEFORES. Between 55–100 s, PEGSNet underestimates by ~0.3 Mw units but captures the increasing trend; it reaches the correct final Mw after ~120 s.
• Alert timing at threshold Mw=8.0: PEGSNet would alert at 53 s after origin, ~8 s earlier than BEFORES. PEGSNet’s error ≤ 0.3 Mw units from 55 s onward; BEFORES reaches similar precision only after ~100 s.
• Robustness: replacing pre-P-wave data with noise yields Mw(t) ≈ 6.5 baseline and never exceeds 8.3, indicating reliance on PEGS. For subduction earthquakes Mw ≥ 7 since 2003 (excluding aftershocks), events with Mw < 8 converge to the noise baseline, confirming lack of detectability below the sensitivity threshold.

The results demonstrate that PEGS contain actionable, instantaneous information about the evolving moment release of large earthquakes, enabling real-time tracking of Mw(t) once the event exceeds ~Mw 8.3. Unlike seismic EEWS that suffer from magnitude saturation and latency tied to P-wave propagation, PEGSNet leverages pre-P-wave recordings to provide zero-delay updates of rupture growth from ~40–55 s after origin for great events. Compared to GNSS-based EEWS, PEGSNet achieves faster and more accurate early estimates without requiring assumptions about slip distributions or station-specific data quality. The method can complement existing seismic and GNSS EEWS to reduce latency and improve accuracy for megathrust earthquakes, and its continuous Mw(t) updates are particularly valuable for tsunami forecasting models that require rapid magnitude evolution. Because PEGS are sensitive to smooth moment-rate variations and propagate effectively at light speed, PEGSNet provides a robust, independent observable to constrain large-event rupture evolution in real time.

The study introduces PEGSNet, a deep learning framework that unlocks true real-time tracking of earthquake growth using prompt elastogravity signals recorded before P-wave arrivals. Trained on extensive synthetic data augmented with real station noise, PEGSNet accurately and instantaneously reconstructs Mw(t) for great earthquakes and outperforms or matches leading EEWS approaches in latency and accuracy during retrospective evaluation of the 2011 Tohoku-Oki event. This approach can immediately enhance tsunami early warning and complement existing EEWS by providing rapid, saturation-free magnitude evolution. PEGSNet is adaptable to other regions and networks with minimal modifications, primarily requiring noise recordings for the target network. Future work could lower sensitivity thresholds via improved noise handling, incorporate more complex source models (finite faults), integrate joint PEGS-GNSS/seismic inputs, and deploy real-time systems in other subduction zones.

• Sensitivity threshold: reliable detection and tracking begin around Mw ≈ 8.3 (improvable to ~7.9–8.0 under low-noise networks), limiting applicability to the largest events.
• Early-time underestimation: the first 30–40 s after origin are underestimated due to threshold effects and potential P-wave masking at near-source stations; at least ~15 s are needed even in noise-free conditions.
• Dependence on P-wave arrival picks: the method assumes known P-wave arrival times at stations to zero post-P portions; practical real-time performance depends on robust triggering.
• Source simplification: training uses point-source approximations; while differences due to source finiteness may be within uncertainties for PEGS, real ruptures are finite and complex. STFs are randomized to mimic empirical behavior but may not capture all complexities.
• Network and preprocessing: performance is noise-dependent and may introduce some latency from preprocessing; requires dense broadband networks and availability of noise recordings for new regions.
• Regional training: the model is tailored to the Japanese subduction zone; portability requires retraining or adaptation to other regions’ geometries and noise characteristics.